[LLVMdev] How best to time passes using the API instead of opt? Also, memory leaks when trying to do timing in the API.

[Andrew Clinton]

I got it working with this:

     int          argc = 2;
     const char  *argv[2] = {"myopt", "-time-passes"};

     cl::ParseCommandLineOptions(argc, (char **)argv, "my optimizer");

On 06/01/2011 01:19 PM, Michael Ilseman wrote:
> I have some PassManagers, and I would like to output timing data for
> how long each pass takes to execute with a separate application using
> the API, rather than through an llvm tool. Unfortunately, I'm having
> trouble seeing how to use the existing facilities without using opt.
> Setting llvm::TimePassesIsEnabled before creating the PassManagers
> doesn't seem to output anything, though it is making Timers and
> TimerGroups. Even if I call llvm::getPassTimer for one of the passes
> (and I do get a timer that is initialized), I can't figure out to get
> to its TimerGroup (TG is private, but if I use a debugger, I can get
> it and call printAll successfully). Looking through the code,
> TimerGroup's removeTimer (also called in it's destructor) appears to
> print everything out to the info-output-file (stderr) when its Timers
> are all removed but I see no output, however I'm able to call
> llvm::CreateInfoOutputFile myself, pipe to it, and stuff shows up on
> stderr. In PassManager.cpp, TheTimeInfo (which is a TimingInfo), gets
> assigned once, and as far as I can tell never gets deleted, thus its
> destructor is never called, the Timers are never deleted, and the
> timing data is never output. Valgrind shows these Timers, TimerGroup,
> and TimingInfo as being memory leaks for my application when I enable
> llvm::TimePassesIsEnabled. Is there a way in the API to delete
> TheTimeInfo myself, even though TheTimeInfo is static? Also, if this
> is a bug/deficiency, I'm willing to patch this with some guidance,
> e.g. is putting "delete TheTimingInfo; TheTimeInfo = 0"; in
> PMTopLevelManager's destructor appropriate? Or is there a much easier
> solution that I'm just missing?
>
> Thanks.
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev